AMR-VSF: an adaptive memory correction framework for robust 3D CT aorta segmentation.
Authors
Affiliations (5)
Affiliations (5)
- School of Computer and Artificial Intelligence, Southwest Jiaotong University, Chengdu, 611756, Sichuan, China.
- Pittsburgh Institute, Sichuan University, Chengdu, 610065, Sichuan, China.
- West China Hospital, Sichuan University, Chengdu, 610065, Sichuan, China.
- Pittsburgh Institute, Sichuan University, Chengdu, 610065, Sichuan, China. [email protected].
- School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China. [email protected].
Abstract
Accurate 3D aortic segmentation in CT images is vital for cardiovascular disease diagnosis, surgical planning, and intraoperative navigation. Existing aortic segmentation methods suffer from 3D incoherence (frame-level models) or cumulative memory drift (memory-based video models), failing to meet real-time clinical demands for surgical planning. This study developed the Adaptive Memory Rectification Video Segmentation Framework (AMR-VSF) to integrate both models' strengths, optimizing single-frame anatomical accuracy and cross-frame temporal stability while meeting real-time demands. AMR-VSF uses a 4-stage modular architecture: (1) Frame-level models generate anatomically complete initial masks; (2) MedSAM2 enables cross-frame mask propagation for temporal consistency; (3) Dual "coverage-diameter deviation rate" metrics detect drift, with segmented memory reset (partial clearing for stable areas, full clearing for complex ones) to correct it; (4) Dynamic weight fusion balances local accuracy and global consistency. TensorRT optimized MedSAM2, and the framework was validated on 92 patients' CT aortic data (≈76,000 slices). AMR-VSF (nnUNet-based) achieved optimal performance: Dice Similarity Coefficient (DSC) 0.963 (5.1% higher than MedSAM2 alone), 95% Hausdorff Distance (HD95) 3.32 mm (38.2% lower), and Average Surface Distance (ASD) 0.61 mm (38.4% lower than the latter). TensorRT cut MedSAM2's single-frame inference time from 73.1 to 31.2 ms with no accuracy loss. Its outputs enabled 3D reconstruction of the aorta (complete "aortic arch-thoracic aorta-abdominal aorta-branch vessels" topological connection, anatomy-matched smooth edges), and the framework showed strong compatibility with diverse models, with potential extension to other long-sequence luminal structure segmentation. AMR-VSF addresses existing limitations by combining frame-level and video models, achieving high accuracy and stability. TensorRT enables real-time use, and its 3D models support tasks like aneurysm measurement. It is compatible and extensible, supporting precise cardiovascular care.